Pith. sign in

REVIEW 4 major objections 6 minor 34 references

CEA-LIST at CheckThat! 2025: Evaluating LLMs as Detectors of Bias and Opinion in Text

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Few-shot-prompted LLMs can match or beat fine-tuned smaller language models on multilingual subjectivity detection, with top ranks in Arabic, Polish, Italian, English, German, and multilingual tracks.

desk verdict A useful shared-task report with clean dev experiments, but the central LLM-vs-SLM claim only holds on the English dev set; the leaderboard wins come from an LLM+SLM ensemble. read the letter →

arxiv 2507.07539 v1 pith:42RPWUNA submitted 2025-07-10 cs.CL cs.AI

Akram Elbouanani , Evan Dufraisse , Aboubacar Tuo , Adrian Popescu This is my paper
classification cs.CLcs.AI
keywords largelanguagemodelsfew-shotlearningpromptengineeringsubjectivitydetectionmultilingualNLPin-contextnoisylabelsmacro-F1
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that large language models (LLMs) used with a well-designed few-shot prompt can match or outperform fine-tuned smaller language models (SLMs) for subjectivity detection—deciding whether a sentence states a fact or expresses opinion, evaluation, or bias. The authors support the claim with official results from a multilingual shared task in which their system placed first in Arabic and Polish, second in Italian, third in English and in the multilingual track, and fourth in German. Their most striking evidence is Arabic, where the few-shot LLM beat the second-ranked system by 0.10 macro-F1 while fine-tuned models plateaued; they argue that the Arabic benchmark's annotation inconsistencies hurt label-dependent SLMs more than prompt-based LLMs. If correct, the finding makes prompt-only LLM pipelines a practical alternative to fine-tuning when labeled data are scarce, unbalanced, or noisily annotated.

What carries the argument

The load-bearing mechanism is the extended few-shot prompt: a detailed natural-language specification of the subjectivity criteria taken from the annotation guidelines (including edge cases for quotes and reported speech), combined with six in-context examples chosen at random with equal class balance. This prompt fixes the under-specification that makes a bare prompt over-predict subjectivity, raising English macro-F1 from 0.54 to 0.76 when examples are added, and it transfers across languages through translation. Random example selection consistently matched or beat similarity-based and dissimilarity-based selection, and the debate-based multi-agent variant added at most one point of macro-F1, so the prompt-plus-examples combination is what does the work.

What would settle it

Run the exact translated six-shot extended-prompt system on every language with per-language validation and compare macro-F1 against fine-tuned transformers; if the LLM advantage shrinks or reverses in languages that never informed prompt design, or if reannotating a sample of Arabic sentences removes the 0.10 macro-F1 gap, the central claim would be refuted.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a single recipe—an extended prompt written from the official annotation guidelines, six randomly selected balanced in-context examples, and an LLM rather than a fine-tuned transformer—is enough to be highly competitive across languages, and that this recipe is especially effective where training labels are noisy. The paper reports that this system ranked first in Arabic and Polish, second in Italian, third in English and multilingual, and fourth in German in the competition's official scoring, and that on Arabic it outperformed the runner-up by 0.10 macro-F1. It further claims that fancy elaborations, such as multi-agent debate or similarity-based example selection, give little beyond the standard few-shot prompt, so the prompt itself, not the complexity of the scaffolding, is what carries the performance. The authors attribute the Arabic result to tolerance of annotation inconsistency, citing the fact that fine-tuned models and even their own LLM variants plateau below 0.55 macro-F1 on that language, while the same pipeline reaches 0.84 on the well-annotated 2023 Arabic data.

Load-bearing premise

The paper assumes the prompt and few-shot choices tuned on English and Arabic transfer to Italian, German, Bulgarian, Romanian, Polish, Ukrainian, and Greek without per-language validation.

Editorial extensions

If this is right

  • A prompt-based LLM pipeline can replace fine-tuned transformers in multilingual subjectivity detection without per-language training data, provided a translated guideline-style prompt and balanced random examples are used.
  • In low-resource or inconsistently labeled settings, LLM few-shot prompting is a stronger default than fine-tuning an SLM on the noisy training set.
  • Annotation quality, not linguistic difficulty or model capacity, can determine leaderboard outcomes; Arabic's low scores across models suggest benchmark noise rather than a hard task.
  • Expensive prompt-engineering variants—debate, adversarial reasoning, similarity sampling—do not justify their cost over a carefully phrased standard few-shot prompt.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's noise-resilience explanation is untested as a general mechanism; one direct extension would be to inject controlled label noise into clean datasets and check whether the LLM few-shot advantage grows with noise level.
  • If the Arabic annotation-inconsistency finding transfers, reannotating that test set could change the official ranking and would provide a cleaner benchmark for future subjectivity systems.
  • The English-to-Arabic transfer assumption is stated but not tested per language; a per-language validation of the translated prompt could show where the one-recipe-fits-all approach starts to fail.
  • The random-beats-similarity result, if it generalizes, would cut the cost of building subjectivity detectors by removing the need for embedding-based example retrieval.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This paper describes the CEA-LIST submission to the CheckThat! 2025 Task 1 (subjectivity detection). The authors experiment with prompt engineering, few-shot in-context learning, debating-style multi-agent setups, and fine-tuned smaller language models (SLMs) on dev data, and then report official test-set results. The central claim is that LLMs with carefully designed few-shot prompts can match or outperform fine-tuned SLMs, particularly under noisy or inconsistent annotations. The paper places first in Arabic and Polish, second in Italian, and third in English and multilingual tracks, and it attributes the strong Arabic result to the method's resilience to annotation noise. The paper also reports a dataset-quality investigation for Arabic, including manual reannotation and translation experiments.

Significance. If the central claim were fully supported, the paper would offer a practical result for the shared-task community: prompt-based LLMs as a viable, low-resource alternative to fine-tuned SLMs for multilingual subjectivity detection. The paper's strengths include the concrete official evaluation against an external benchmark, the detailed prompt templates in the appendix, and the explicit investigation of Arabic annotation quality. However, the evidence as presented does not isolate the LLM contribution from the ensemble, and the only controlled LLM-only versus SLM-only comparison is a single English dev set with no statistical significance assessment. The paper is a useful system description, but the abstract's broad claim is currently overreaching.

major comments (4)
  1. [Section 5.2.1 and Table 8] The official leaderboard results were produced by an ensemble that includes the fine-tuned RoBERTa-Base SLM alongside GPT-4 variants, LLaMA 70B, and Qwen 72B, yet the abstract states that 'LLMs, when paired with carefully designed prompts, can match or outperform fine-tuned smaller language models.' The leaderboard rankings therefore cannot directly support that claim, since they do not isolate the LLM contribution. The authors should provide a per-language ablation separating LLM-only, SLM-only, and ensemble predictions on the official test sets, or revise the claim to describe the system as an ensemble rather than as LLMs alone.
  2. [Section 5.1.3 and Table 4] The text says that 'random selection outperforms similarity-based strategies across all models,' but Table 4 shows LLaMA 70B with dissimilarity-based selection at 0.75 Macro F1 versus random at 0.73, and Qwen 72B with dissimilarity at 0.73 versus random at 0.71. The claim is also internally inconsistent with the following sentence about Qwen, which says dissimilar sampling 'offers minimal F1 gain' while the table shows a gain over both random and similarity. The reported results need to be reconciled with the text, and the comparison should separate random-vs-similarity from random-vs-dissimilarity.
  3. [Tables 3-7 and Section 5.2.1] All dev-set experiments appear to be single runs with no variance or significance testing, and differences of 0.01-0.02 in Macro F1 (e.g., 0.76 vs 0.77 in Tables 3 and 6) are used to conclude that advanced strategies 'did not lead to substantial improvements' or that example selection had 'no significant impact.' For a claim of equivalence, the authors should report multiple random seeds (especially for random few-shot selection), standard deviations, or pairwise significance tests, or at least hedge the conclusions as observational rather than statistically established.
  4. [Section 3 and Section 5.2.2] The paper explicitly assumes that insights from English and Arabic transfer to the other languages, and the final system applies the same English-derived prompt recipe to Italian, German, Bulgarian, and the zero-shot languages without per-language validation. The conclusion nonetheless emphasizes 'diverse languages and settings' and multilingual robustness. To support this, the authors should report dev-set results for each language in the LLM-only few-shot setup (and, where feasible, for the ensemble) so that the transferability assumption is empirically checked rather than postulated.
minor comments (6)
  1. [Section 5.1.3] The sentence 'Similarity is measured using the cosine similarity between sentence embeddings generated by GPT-3' contradicts Section 4.2, which specifies OpenAI's text-embedding-3-small; please make the embedding model reference consistent.
  2. [Section 6] The example sentences following 'the sentence:' appear to be missing from the manuscript, so the reader cannot verify the claimed annotation inconsistencies; please insert the Arabic sentences or refer to a supplementary file.
  3. [Section 6] The manual reannotation by a single native Arabic speaker is reported without inter-annotator agreement or a clear procedure; it should be described as a preliminary, anecdotal check rather than a quantitative validation of annotation noise.
  4. [Section 5.2.1] The list of models in the official ensemble ('GPT-4 variants (GPT-4o-mini, GPT-4.1-mini), RoBERTa, LLaMA 70B, and Qwen 72B') does not match Table 7's 'LLM Ensemble' list, which includes Aya-Expanse 32B; please clarify which models were actually used in the official submission.
  5. [Section 5.2.2] The text claims 'top-three positions in the majority of the evaluated languages,' but Table 8 shows rank 4 in German, 6 in Romanian, 7 in Greek, and 10 in Ukrainian; please adjust the wording to be consistent with the table.
  6. [Section 6] The reported 2023 Arabic result of F1 = 0.84 is presented without specifying the model, prompt, or few-shot configuration; please provide the necessary details for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical evaluation against an external benchmark, and no load-bearing step reduces to its own inputs or to self-citation.

full rationale

The paper reports a shared-task submission to CheckThat! 2025 and evaluates LLM-based few-shot prompting against an external test set. The central claim—that LLMs with carefully designed prompts can match or outperform fine-tuned SLMs—is supported by official leaderboard scores (Table 8) and by controlled experiments on the English development set (Tables 2–7). No parameter is fitted to the test labels and then reported as a prediction; few-shot examples are drawn from the provided training data, and the prompts are constructed from the task's annotation guidelines. The Arabic noise-robustness explanation (Section 5.2.2) is a post-hoc attribution, not a quantity derived from the method itself, and Section 6 explicitly reports the underlying reproducibility gap. The acknowledged assumption in Section 3 that insights transfer across languages is a stated limitation, not a circular step. The concern that the official system was an ensemble including RoBERTa (Section 5.2.1), so the leaderboard does not isolate LLM-only performance, is an evidence-attribution issue rather than a circularity issue: the scores are still externally measured, and the paper does not claim to derive those scores from a fitted quantity. There are no load-bearing self-citations, no imported uniqueness theorems, and no renaming of a known result. The derivation chain is therefore self-contained with respect to circularity, and the correct score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

This is an empirical system paper, so the ledger records the design choices that were tuned on the development set and the stated assumptions about cross-lingual transfer. No free parameters are fitted to a target quantity in a derivation sense; the 'free parameters' are hyperparameters selected by performance on the dev set. There are no invented entities.

free parameters (3)
  • Few-shot example count = 6
    The 6-shot setting was chosen after comparing 0, 6, and 12-shot prompts on the development set (Table 3); 6-shot and 12-shot gave equal macro F1 (0.76) and 6-shot was selected for the final system.
  • Example selection strategy = random
    Random selection outperformed similarity and dissimilarity selection in Table 4 and was adopted for the final submission.
  • Prompt variant = extended prompt
    The extended prompt derived from annotation guidelines outperformed the basic prompt (Table 3) and was used in the final system.
assumptions (3)
  • domain assumption Insights from English and Arabic experiments transfer to other languages
    Section 3 states: 'We operate under the assumption that insights obtained from these languages are transferable to the other languages.' The final system applies the English-tuned prompt design to all languages without per-language validation except the official test set.
  • domain assumption DeepL translations of the extended prompt are semantically equivalent across languages
    Appendix A.2 notes prompts for other languages were translated with DeepL; no human verification of translation quality is reported.
  • domain assumption The official CheckThat! 2025 test labels are ground truth
    All final results (Table 8) are taken from the official evaluation leaderboard; the paper does not question these labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CEA-LIST at CheckThat! 2025: Evaluating LLMs as Detectors of Bias and Opinion in Text." pith.science (2026). https://pith.science/paper/42RPWUNA

@misc{pith2026250707539,
  author       = {Pith},
  title        = {Pith review of: CEA-LIST at CheckThat! 2025: Evaluating LLMs as Detectors of Bias and Opinion in Text},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42RPWUNA}},
  note         = {Machine review of arXiv:2507.07539}
}
read the original abstract

This paper presents a competitive approach to multilingual subjectivity detection using large language models (LLMs) with few-shot prompting. We participated in Task 1: Subjectivity of the CheckThat! 2025 evaluation campaign. We show that LLMs, when paired with carefully designed prompts, can match or outperform fine-tuned smaller language models (SLMs), particularly in noisy or low-quality data settings. Despite experimenting with advanced prompt engineering techniques, such as debating LLMs and various example selection strategies, we found limited benefit beyond well-crafted standard few-shot prompts. Our system achieved top rankings across multiple languages in the CheckThat! 2025 subjectivity detection task, including first place in Arabic and Polish, and top-four finishes in Italian, English, German, and multilingual tracks. Notably, our method proved especially robust on the Arabic dataset, likely due to its resilience to annotation inconsistencies. These findings highlight the effectiveness and adaptability of LLM-based few-shot learning for multilingual sentiment tasks, offering a strong alternative to traditional fine-tuning, particularly when labeled data is scarce or inconsistent.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 25 canonical work pages

  1. [1]

    Ruggeri, F

    F. Ruggeri, F. Antici, A. Galassi, A. Korre, A. Muti, A. Barron, et al., On the definition of pre- scriptive annotation guidelines for language-agnostic subjectivity detection, in: CEUR Workshop Proceedings, volume 3370, CEUR-WS, 2023, pp. 103–111

    work page 2023

  2. [2]

    H. Yu, V. Hatzivassiloglou, Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences, in: Proceedings of the 2003 conference on Empirical methods in natural language processing, 2003, pp. 129–136

    work page 2003

  3. [3]

    Abimbola, E

    B. Abimbola, E. de La Cal Marin, Q. Tan, Enhancing legal sentiment analysis: A convolu- tional neural network–long short-term memory document-level model, Machine Learning and Knowledge Extraction 6 (2024) 877–897. URL: https://www.mdpi.com/2504-4990/6/2/41. doi:10.3390/make6020041

    work page doi:10.3390/make6020041 2024

  4. [4]

    Liu, Sentiment analysis and opinion mining, Springer Nature, 2022

    B. Liu, Sentiment analysis and opinion mining, Springer Nature, 2022

    work page 2022

  5. [5]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al., Language models are few-shot learners, Advances in neural information processing systems 33 (2020) 1877–1901

    2020

  6. [6]

    Galassi, F

    A. Galassi, F. Ruggeri, A. Barrón-Cedeño, F. Alam, T. Caselli, M. Kutlu, J. M. Struß, F. Antici, M. Hasanain, J. Köhler, et al., Overview of the clef-2023 checkthat! lab: Task 2 on subjectivity in news articles, in: 24th Working Notes of the Conference and Labs of the Evaluation Forum, CLEF-WN 2023, CEUR Workshop Proceedings (CEUR-WS. org), 2023, pp. 236–249

    work page 2023

  7. [7]

    J. M. Struß, F. Ruggeri, A. Barrón-Cedeño, F. Alam, D. Dimitrov, A. Galassi, G. Pachov, I. Koychev, P. Nakov, M. Siegel, et al., Overview of the clef-2024 checkthat! lab task 2 on subjectivity in news articles, in: CEUR Workshop Proceedings, volume 3740, CEUR-WS, 2024, pp. 287–298

    work page 2024

  8. [8]

    A. I. Paran, M. S. Hossain, S. H. Shohan, J. Hossain, S. Ahsan, M. M. Hoque, Semanticcuetsync at checkthat! 2024: finding subjectivity in news article using llama, Faggioli et al.[22] (2024)

    work page 2024

Show all 34 references
  1. [9]

    Javdan, B

    S. Javdan, B. Minaei-Bidgoli, et al., Applying transformers and aspect-based sentiment analysis approaches on sarcasm detection, in: Proceedings of the second workshop on figurative language processing, 2020, pp. 67–71

    2020

  2. [10]

    Zhang, Y

    W. Zhang, Y. Deng, B. Liu, S. Pan, L. Bing, Sentiment analysis in the era of large language models: A reality check, in: K. Duh, H. Gomez, S. Bethard (Eds.), Findings of the Association for Computational Linguistics: NAACL 2024, Association for Computational Linguistics, Mexic...

    2024

  3. [11]

    Shokri, V

    M. Shokri, V. Sharma, E. Filatova, S. Jain, S. Levitan, Subjectivity detection in english news using large language models, in: Proceedings of the 14th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, 2024, pp. 215–226

    2024

  4. [12]

    Huang, E

    T. Huang, E. Fan, Structured reasoning for fairness: A multi-agent approach to bias detection in textual data, 2025. URL: https://arxiv.org/abs/2503.00355. arXiv:2503.00355

    2025 arXiv

  5. [13]

    Suwaileh, M

    R. Suwaileh, M. Hasanain, F. Hubail, W. Zaghouani, F. Alam, Thatiar: subjectivity detection in arabic news sentences, arXiv preprint arXiv:2406.05559 (2024)

    2024 arXiv

  6. [14]

    Kusrini, M. Mashuri, Sentiment analysis in twitter using lexicon based and polarity multiplication, in: 2019 International Conference of Artificial Intelligence and Information Technology (ICAIIT), 2019, pp. 365–368. doi:10.1109/ICAIIT.2019.8834477

    2019

  7. [15]

    Zahoor, R

    S. Zahoor, R. Rohilla, Twitter sentiment analysis using lexical or rule based approach: A case study, in: 2020 8th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions) (ICRITO), 2020, pp. 537–542. doi:10.1109/ICRITO48877...

    2020

  8. [16]

    Kotelnikova, D

    A. Kotelnikova, D. Paschenko, K. Bochenina, E. Kotelnikov, Lexicon-based methods vs. bert for text sentiment analysis, in: International Conference on Analysis of Images, Social Networks and Texts, Springer, 2021, pp. 71–83

    2021

  9. [17]

    R. A. Potamias, G. Siolas, A.-G. Stafylopatis, A transformer-based approach to irony and sarcasm detection, Neural Computing and Applications 32 (2020) 17309–17320

    2020

  10. [18]

    Marvin, N

    G. Marvin, N. Hellen, D. Jjingo, J. Nakatumba-Nabende, Prompt engineering in large language models, in: International conference on data intelligence and cognitive informatics, Springer, 2023, pp. 387–402

    2023

  11. [19]

    Sahoo, A

    P. Sahoo, A. K. Singh, S. Saha, V. Jain, S. Mondal, A. Chadha, A systematic survey of prompt engi- neering in large language models: Techniques and applications, arXiv preprint arXiv:2402.07927 (2024)

    2024 arXiv

  12. [20]

    J. Zhuo, S. Zhang, X. Fang, H. Duan, D. Lin, K. Chen, Prosa: Assessing and understanding the prompt sensitivity of llms, 2024. URL: https://arxiv.org/abs/2410.12405. arXiv:2410.12405

    2024 arXiv

  13. [21]

    Errica, G

    F. Errica, G. Siracusano, D. Sanvito, R. Bifulco, What did i do wrong? quantifying llms’ sen- sitivity and consistency to prompt engineering, 2025. URL: https://arxiv.org/abs/2406.12334. arXiv:2406.12334

    2025 arXiv

  14. [22]

    Y. Wang, Q. Yao, J. T. Kwok, L. M. Ni, Generalizing from a few examples: A survey on few-shot learning, ACM computing surveys (csur) 53 (2020) 1–34

    2020

  15. [23]

    Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, I. Mordatch, Improving factuality and reasoning in language models through multiagent debate, in: Forty-first International Conference on Machine Learning, 2024. URL: https://openreview.net/forum?id=zj7YuTE4t8

    2024

  16. [24]

    Liang, Z

    T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, S. Shi, Z. Tu, Encouraging divergent thinking in large language models through multi-agent debate, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural ...

    2024 doi

  17. [25]

    Shinn, F

    N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, S. Yao, Reflexion: Language agents with verbal reinforcement learning, Advances in Neural Information Processing Systems 36 (2023) 8634–8652

    2023

  18. [26]

    F. Alam, J. M. Struß, T. Chakraborty, S. Dietze, S. Hafid, K. Korre, A. Muti, P. Nakov, F. Ruggeri, S. Schellhammer, V. Setty, M. Sundriyal, K. Todorov, V. V., The clef-2025 checkthat! lab: Subjectivity, fact-checking, claim normalization, and retrieval, in: C. Hauff, C. Macdo...

    2025

  19. [27]

    F. Alam, J. M. Struß, T. Chakraborty, S. Dietze, S. Hafid, K. Korre, A. Muti, P. Nakov, F. Ruggeri, S. Schellhammer, V. Setty, M. Sundriyal, K. Todorov, V. Venktesh, Overview of the CLEF-2025 CheckThat! Lab: Subjectivity, fact-checking, claim normalization, and retrieval, in: ...

    2025

  20. [28]

    Atanasova, A

    P. Atanasova, A. Barron-Cedeno, T. Elsayed, R. Suwaileh, W. Zaghouani, S. Kyuchukov, G. D. S. Martino, P. Nakov, Overview of the clef-2018 checkthat! lab on automatic identification and verification of political claims. task 1: Check-worthiness, arXiv preprint arXiv:1808.05542 (2018)

    2018 arXiv

  21. [29]

    Atanasova, P

    P. Atanasova, P. Nakov, G. Karadzhov, M. Mohtarami, G. Da San Martino, Overview of the clef-2019 checkthat! lab: Automatic identification and verification of claims. task 1: Check-worthiness., CLEF (Working Notes) 2380 (2019)

    2019

  22. [30]

    Shaar, A

    S. Shaar, A. Nikolov, N. Babulkov, F. Alam, A. Barrón-Cedeno, T. Elsayed, M. Hasanain, R. Suwaileh, F. Haouari, G. Da San Martino, et al., Overview of checkthat! 2020 english: Automatic identification and verification of claims in social media., CLEF (working notes) 2696 (2020)

    2020

  23. [31]

    Shaar, M

    S. Shaar, M. Hasanain, B. Hamdan, Z. S. Ali, F. Haouari, A. Nikolov, M. Kutlu, Y. S. Kartal, F. Alam, G. Da San Martino, et al., Overview of the clef-2021 checkthat! lab task 1 on check-worthiness estimation in tweets and political debates., in: CLEF (working notes), 2021, pp. 369–392

    2021

  24. [32]

    B. Xu, Q. Wang, Z. Mao, Y. Lyu, Q. She, Y. Zhang, 𝑘nn prompting: Beyond-context learning with calibration-free nearest neighbor inference, 2023. URL: https://arxiv.org/abs/2303.13824. arXiv:2303.13824

    2023 arXiv

  25. [33]

    S. Min, X. Lyu, A. Holtzman, M. Artetxe, M. Lewis, H. Hajishirzi, L. Zettlemoyer, Rethinking the role of demonstrations: What makes in-context learning work?, in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Lang...

    2022 doi

  26. [34]

    so damaged

    K. Dey, P. Tarannum, M. A. Hasan, S. R. H. Noori, Nn at checkthat!-2023: Subjectivity in news articles classification with transformer based models., in: CLEF (Working Notes), 2023, pp. 318–328. A. Prompts Used In this section, we report the prompts used for the classification...

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.